Year: 2021
Type: conference-abstract
Abstract: Rational: Following recommendations from the United States Preventive Service Task Force (USPSTF), low dose computed tomography (LDCT) became the primary screening modality for individuals at high-risk of lung cancer, reducing the mortality rate by 20%. However, the adoption of LDCT screening has led to an increase in the detection of indeterminate nodules, exposing patients to invasive procedures for what is often a benign nodule. Current clinical practice fails to take into consideration risk factors for malignancy that are reflected in local lung parenchyma, such as emphysema. In this study, we use the National Lung Screening Trial (NLST) to investigate how the accuracy of malignancy detection is affected by the attention of a convolutional neural network (CNN) on the local parenchyma versus focusing solely on the nodule. Methods: Selecting nodules between 4mm and 20mm in diameter from the NLST LDCT data set, a CNN was used to explore the performance of parenchymal features for nodule classification [1]. Regions of interest (ROI) containing the nodules were used to train the CNN to extract the morphological characteristics of both parenchyma and nodule (Original). A secondary dataset (Obfuscated) was generated by masking the nodules within each ROI; an Otsu algorithm was used to differentiate between high and low intensity pixels in order to obscure the nodule. A tertiary dataset (Combined) was created by combining the original and obfuscated dataset. Five-fold cross validation, with 5 repetitions, were used to train the model. Results: Applying the Original, Obfuscated, and Combined datasets to our architecture resulted in a mean area under the operator-receiver curve (AUC) of 92.67 (+/-2.4) %, 96.31 (+/-0.9) %, 93.4 (+/-1.7) %, respectively (Figure The Obfuscated dataset performed significantly better compared to either the Original or Combined dataset with adjusted p-values of 0.019 and 0.004, respectively. Conclusion: Our results suggest that masking nodules and other high intensity signals in an image allows a CNN to learn less dominant information present within the local parenchyma. However, because masking only hides the intensity and texture, the still discernible shapes and sizes of the nodules may be contributing to the higher classification accuracy of the obfuscated dataset. Nevertheless, our results indicate that the surrounding parenchyma contains critical information related to a diagnosis of nodule malignancy, which can be readily extracted via deep leaning.
Institution: University of Vermont
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4238075310